| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | |
| 5 | from http import HTTPStatus |
| 6 | |
| 7 | from fastapi import APIRouter, Depends, FastAPI, Request |
| 8 | from fastapi.responses import JSONResponse, StreamingResponse |
| 9 | |
| 10 | from vllm.entrypoints.anthropic.protocol import ( |
| 11 | AnthropicCountTokensRequest, |
| 12 | AnthropicCountTokensResponse, |
| 13 | AnthropicError, |
| 14 | AnthropicErrorResponse, |
| 15 | AnthropicMessagesRequest, |
| 16 | AnthropicMessagesResponse, |
| 17 | ) |
| 18 | from vllm.entrypoints.anthropic.serving import AnthropicServingMessages |
| 19 | from vllm.entrypoints.serve.engine.protocol import ErrorResponse |
| 20 | from vllm.entrypoints.serve.exception_handling.error_response import ( |
| 21 | create_error_response, |
| 22 | ) |
| 23 | from vllm.entrypoints.serve.utils.api_utils import ( |
| 24 | load_aware_call, |
| 25 | validate_json_request, |
| 26 | with_cancellation, |
| 27 | ) |
| 28 | from vllm.logger import init_logger |
| 29 | |
| 30 | logger = init_logger(__name__) |
| 31 | |
| 32 | router = APIRouter() |
| 33 | |
| 34 | |
| 35 | def messages(request: Request) -> AnthropicServingMessages: |
| 36 | return request.app.state.anthropic_serving_messages |
| 37 | |
| 38 | |
| 39 | def translate_error_response(response: ErrorResponse) -> JSONResponse: |
| 40 | anthropic_error = AnthropicErrorResponse( |
| 41 | error=AnthropicError( |
| 42 | type=response.error.type, |
| 43 | message=response.error.message, |
| 44 | ) |
| 45 | ) |
| 46 | return JSONResponse( |
| 47 | status_code=response.error.code, content=anthropic_error.model_dump() |
| 48 | ) |
| 49 | |
| 50 | |
| 51 | @router.post( |
| 52 | "/v1/messages", |
| 53 | dependencies=[Depends(validate_json_request)], |
| 54 | responses={ |
| 55 | HTTPStatus.OK.value: {"content": {"text/event-stream": {}}}, |
| 56 | HTTPStatus.BAD_REQUEST.value: {"model": AnthropicErrorResponse}, |
| 57 | HTTPStatus.NOT_FOUND.value: {"model": AnthropicErrorResponse}, |
| 58 | HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": AnthropicErrorResponse}, |
| 59 | }, |
| 60 | ) |
| 61 | @with_cancellation |
| 62 | @load_aware_call |
| 63 | async def create_messages(request: AnthropicMessagesRequest, raw_request: Request): |
| 64 | handler = messages(raw_request) |
| 65 | if handler is None: |
| 66 | base_server = raw_request.app.state.serving_tokenization |
| 67 | error = base_server.create_error_response( |
| 68 | NotImplementedError("The model does not support Messages API") |
| 69 | ) |
| 70 | return translate_error_response(error) |
| 71 | |
| 72 | try: |
| 73 | generator = await handler.create_messages(request, raw_request) |
| 74 | except Exception as e: |
| 75 | logger.exception("Error in create_messages: %s", e) |
| 76 | return translate_error_response(create_error_response(e)) |
| 77 | |
| 78 | if isinstance(generator, ErrorResponse): |
| 79 | return translate_error_response(generator) |
| 80 | |
| 81 | elif isinstance(generator, AnthropicMessagesResponse): |
| 82 | resp = generator.model_dump(exclude_none=True) |
| 83 | logger.debug("Anthropic Messages Response: %s", resp) |
| 84 | return JSONResponse(content=resp) |
| 85 | |
| 86 | return StreamingResponse(content=generator, media_type="text/event-stream") |
| 87 | |
| 88 | |
| 89 | @router.post( |
| 90 | "/v1/messages/count_tokens", |
| 91 | dependencies=[Depends(validate_json_request)], |
| 92 | responses={ |
| 93 | HTTPStatus.OK.value: {"model": AnthropicCountTokensResponse}, |
| 94 | HTTPStatus.BAD_REQUEST.value: {"model": AnthropicErrorResponse}, |
| 95 | HTTPStatus.NOT_FOUND.value: {"model": AnthropicErrorResponse}, |
| 96 | HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": AnthropicErrorResponse}, |
| 97 | }, |
| 98 | ) |
| 99 | @with_cancellation |
| 100 | @load_aware_call |
| 101 | async def count_tokens(request: AnthropicCountTokensRequest, raw_request: Request): |
| 102 | handler = messages(raw_request) |
| 103 | if handler is None: |
| 104 | base_server = raw_request.app.state.serving_tokenization |
| 105 | error = base_server.create_error_response( |
| 106 | NotImplementedError("The model does not support Messages API") |
| 107 | ) |
| 108 | return translate_error_response(error) |
| 109 | |
| 110 | try: |
| 111 | response = await handler.count_tokens(request, raw_request) |
| 112 | except Exception as e: |
| 113 | logger.exception("Error in count_tokens: %s", e) |
| 114 | return translate_error_response(create_error_response(e)) |
| 115 | |
| 116 | if isinstance(response, ErrorResponse): |
| 117 | return translate_error_response(response) |
| 118 | |
| 119 | return JSONResponse(content=response.model_dump(exclude_none=True)) |
| 120 | |
| 121 | |
| 122 | def attach_router(app: FastAPI): |
| 123 | app.include_router(router) |